No weights after training

[K-R-O-K-O]

Hey everyone, I am trying to develop my own CNN for diagnostic application. I constructed the custom CNN and dataloader, did the training successfully. In log dir I obtained checkpoint.pth.tar, best.pth.tar, qat_best.pth.tar, qat_checkpoint.pth.tar. Now I copied one of this files to ai8x-synthetizer to convert it to .c files but I'm obtaining WARNING: All weights for conv9.op.weightare zero. for every layer which is not right.

Am I missing some step between the training and the synthetizer phase?

[ermanok]

First of all, we suggest using QAT models to have better performance with quantized model. All the models in the log directory are floating point but qat models are trained to optimize the quantized weights. Therefore, whichever model you want to use, you have to run quantize.py to quantize your floating point model. I doubt you skip this quantization step before synthesize the model for the hardware.